OCELOT: Odometry and Contact Estimation for Legged Robots

Emre Girgin, Cagri Kilic · 2026

Read OCELOT: Odometry and Contact Estimation for Legged Robots by Emre Girgin, Cagri Kilic (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-21863.txt.

Introduction

Emre Girgin and Cagri Kilic present OCELOT, leg odometry for legged robots using only onboard sensors. Feet in stance correct an error-state extended Kalman filter state, fed by an inertial unit, encoders and force sensors. The core fuses contact detection and uncertainty to reject slippage with two detectors per foot. They are a force-based Gaussian mixture model finite state machine and a kinematic generalized likelihood ratio test. On 29 sequences, 2.4 kilometers of terrain, it handles slippery environments, with code and a real-time open-source ROS2 package.

Catalog record

CREATOREmre Girgin, Cagri Kilic
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38200
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.21863

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

OCELOT: Odometry and Contact Estimation for Legged Robots Emre Girgin, Cagri Kilic Abstract One of the significant challenges in legged robotics is achieving accurate odometry using only onboard proprioceptive sensors. In this study, we present a complete leg odometry pipeline based on an Error-State EKF (ESEKF) that relies exclusively on proprioceptive data: a body fixed IMU, joint encoders, and force sensors, where filter’s state is corrected by feet determined to be in a stationary stance. The core of our contribution is fused contact detection and an uncertainty quantification module designed to explicitly identify and reject slippage. This module runs two detectors in parallel for each foot, 1) a debounced, force-based Gaussian Mixture Model (GMM) guided Finite State Machine (FSM) to confirm physical contact, and 2) a kinematic-based Generalized Likelihood Ratio Test (GLRT) on the estimated velocity of the foot. The continuous quality scores from both estimators are fused to detect if the foot is both physically loaded and kinematically stationary and served as an uncertainty signal for each contact. To validate our approach, we collected a multi-modal dataset of 29 sequences spanning diverse indoor and outdoor terrains (e.g., concrete, grass, pebble, and rock) total of 2.4 km long. We benchmarked our approach against both proprioceptive and exteroceptive methods. The results demonstrate our method’s efficacy in providing accurate odometry estimates, robustly handling slippage-prone environments. We also share our code and real-time ROS2 package as open-source. 1 1 1 Github I INTRODUCTION Legged robots are receiving significant research interest due to their better locomotion capabilities in unstructured and uneven terrains, environments that pose considerable challenges for conventional wheeled robots [ 25 , 8 ] . On the other hand, the absence of wheels necessitates a more complex locomotion strategy, characterized by intermittent contact dynamics as the end-effectors (feet) cyclically make and break contact with the ground [ 7 ] . Tracking the position and velocity of the robot’s main body, or Leg Odometry [ 22 ] , is achieved by exploiting the robot’s kinematic and dynamic models [ 3 ] . Consequently, the fundamental proprioceptive sensor suite for a legged robot typically comprises an IMU and joint encoders to measure angular positions and velocities; this contrasts with the IMU and wheel encoder pairing used on wheeled platforms. While these sensors are standard, the suite is frequently augmented with contact/force sensors to provide explicit information regarding foot-ground interaction. Fig. 1: Our contact estimation method provides robustness against various terrain, including rough granular medium like rock and pebble leading to slippage. The dominant framework in leg odometry is to propagate the robot’s state via an IMU-driven Inertial Navigation System (INS) and correct drift through Zero Velocity Updates (ZUPTs) [ 27 ] when a foot is determined to be in stance. However, accurately applying this correction requires knowing exactly when a leg is stationary, making contact state estimation a critical sub-problem. Existing solutions, reviewed in Section II , range from force-based thresholding to leg-mounted IMUs and learning-based classifiers. In this study, we adopt an ESEKF architecture that uses only the onboard proprioceptive sensor suite of IMU, joint encoders, and foot force sensors. The core of our contribution lies in fused contact detection and uncertainty quantification module. This module runs two detectors in parallel: a force-based FSM which is adaptively governed by a GMM for direct, physical contact, and a kinematic GLRT on estimated foot velocity. By fusing the continuous quality scores from both detectors, our method robustly identifies stance phases which are physically loaded and kinematically stationary. This fused score is then used to dynamically compute an adaptive measurement covariance for each stance based on contact quality. This allows the ESEKF to intelligently weight each measurement regarding its reliability: a high-quality, non-slip contact results in a low measurement covariance and a strong correction, while a low-quality event like a slip results in a high measurement covariance, causing the filter to automatically down-weight its influence. Finally, an innovation based gating step validates each accepted measurement, rejecting any remaining statistical outliers before the final correction. Fig. 2: Overall architecture of the proposed method. Our study utilizes IMU for state propagation and corrects the state based on the contact event for each leg separately. Detecting contact and quantifying the uncertainty is conducted by two distinct detectors and their predictions are fused as the measurement model. We benchmarked our method against both proprioceptive-only methods and exteroceptive approaches and demonstrated our method’s effectiveness. To summarize, our contributions in this paper are: • A novel contact detection and adaptive uncertainty framework that combines a force-based GMM-FSM and a body IMU derived GLRT contact stationarity detector, eliminating the need for supplementary foot-mounted sensors, to generate a continuous contact quality score, which is translated into a dynamic measurement covariance \mathbf{R}_{i} . • An open-source ESEKF-based state estimator for quadruped leg odometry using only onboard proprioceptive sensors (fixed-body IMU, joint encoders, and force sensors). • A new publicly available legged odometry dataset for quadrupeds, including 29 indoor and outdoor sequences collected on diverse terrain types (concrete, tile, grass, pebble and rock) for total of 2.4 km path. II RELATED WORK Proprioceptive leg odometry has been extensively studied using EKFs [ 3 , 2 ] , Invariant EKF (InEKF) [ 4 , 1 ] frameworks [ 13 , 12 , 29 , 11 , 24 , 17 ] , smoothing-based approaches [ 16 , 35 ] , and learning-based alternatives [ 5 , 31 ] . All these various frameworks mainly adopt IMU based propagation and differ by the contact detection method they obtained. Moreover, all these studies highlight a fundamental limitation: the unobservability of absolute position and heading when relying solely on body-fixed IMUs, joint encoders, and force sensors [ 3 , 2 , 29 , 13 ] making proprioceptive leg odometry a challenging problem. II-A Contact Estimation One major category of contact estimation methods relies directly on Ground Reaction Force (GRF) data. Fallon et al. [ 9 ] constructed an FSM based on torque/force measurements to provide debounced contact states, relying on fixed state transitions to determine contact events. Rotella et al. [ 23 ] applied weighted K-means clustering to GRF data to automatically group measurements into distinct contact and swing phases. Conversely, this method assumes the GRF signal space is separable into simple spherical clusters. A second category of methods infers contact without direct force measurements, often relying on robot dynamics or kinematics. Bloesch et al. [ 3 , 2 ] adopts simple thresholding on estimated GRF with static measurement covariance. Camurri et al. [ 6 ] estimated the GRF by formulating a dynamic model that integrated joint angles, motor efforts, and torques generated by external forces using fixed thresholding on the GRF and does not explicitly address foot slippage. Jenelten et al. [ 14 ] proposed a kinematic approach, fitting the vertical component of the estimated foot velocity to an exponential distribution and interpreting peaks in its derivative as touchdown events. While this method rapidly detects slippage, it struggles to consistently classify the complete stance phase. Hartley et al. [ 11 ] estimated the GRF from joint torques and applied logistic regression to adaptively compute contact probability, similar to ours. However, the associated covariance is primarily governed by \Delta GRF, making

Read the complete text